How does Thunderbird “decide” what is junk?
Basically, you tell it. Your question characterizes it as if it's a comprehensive, refined spam filter that you can also add to, but it's really the other way around. It comes with some starter rules, and that's what your question focuses on. Those are pretty generic and imprecise, just a crude baseline to get things going. They aren't intended to be used as-is. I'm not aware of those being published (although Thunderbird is open source, so somebody could always dig through the code). However, its real power comes from your training the adaptive filter.
Thunderbird uses an adaptive filter that learns from your actions which messages are legitimate and which are junk.
In order for this filter to be effective, you must train it to recognize the messages that you consider to be junk and the messages that you consider to be not junk. So you will want to mark messages as junk, not delete them.
You will need to mark many messages so that the adaptive filter has enough training data, including messages that are NOT junk. ... It as just as important to tell the filter which messages are not junk.
First, during the early learning stages you will want to frequently, perhaps daily, check your Junk folder for messages that have been incorrectly classified as junk by clicking on the Not junk button, or using the upper case J on your keyboard. After the first week you should also check the junk folder for messages that are incorrectly marked as junk, perhaps weekly.
Secondly, very important, you will want to constantly train the filter by marking a quantity of GOOD messages as not junk, for example messages in your Inbox. You must use the keyboard upper case J, because there is no button - the "Not Junk" button appears only for messages that have already been classified as junk. Marking several messages per week will be sufficient. You can select many messages and mark them all at the same time. Note - unfortunately nothing in the user interface indicates whether a message has already been marked as "not junk".